Rest of World 跟了 6 位亚非从业者,看他们真拿 AI 干活:搭工具、做生意、做研究,比看厂商发布会实在多了。
Latest
Databricks 分享了自家排障实录:用 Unity Gateway 和 Genie One 追踪 MCP-server,一年白烧 49.9 万美元 token 的 Bug 全揪出来了,做智能体运维的可以抄作业。
Lovart 用 GPT-Image 2.5 一口气生成整套品牌视觉,从 logo 到配色都齐了,做设计的朋友可以看看这套流程。
博主用 Claude Opus 5.5 加一个视频技能包,丢个网址就自动剪出播客宣传片,还顺带聊了 JEV 模型 5 美元免费额度怎么玩。
做 code review 可以试试 astra,多位开发者公认最彻底,能挖出高危漏洞,就是费 token。
两家泰国公司现场聊怎么用 WonderClip 做视频,讲的是工作流和本地化落地的实战经验。
The sessions focus on concrete agent evaluation techniques — tuned evaluators and trajectory-based assessment — that most teams building production agents have not yet standardised. Practitioners evaluating LLM applications can attend the AMAs directly and apply the walkthroughs to their own LangSmith workflows. The 21 October session on deep agents with managed infrastructure signals how LangChain is packaging deployment, not just observability.
OpenClaw 把自家删了 40 万行测试的 skill 开源了,agent 写测试前先过四个问题,做 AI 编程的可以直接抄这个工作流。
歸藏试了 Claude Opus 5.5 一键给产品做宣传片,效果比 GPT-6 Astra 还好,还要把这套玩法蒸馏给便宜模型用。
Running MoE models on commodity ARM hardware rather than GPUs can cut inference costs for edge and on-premise deployments, and this session pairs two inference stacks, vLLM and OpenVINO, with hands-on strategies. Engineers deploying LLMs on ARM servers or devices, including the growing ARM server fleet in Southeast Asian data centres, can pick up concrete optimisation techniques from it.
Vague cleanup prompts make coding agents stop early, while hard numeric constraints push them to finish a real pruning pass. Teams using Claude, GPT or other agents for maintenance can apply the same coverage-window constraint to their own AI-swollen test suites. The metric-first prompting pattern transfers directly to refactoring tasks beyond tests.
有人把参考解说视频丢给 Claude Opus 5.5,换个主题就让它照着风格复刻出片,做视频的可以试试这套流程。
LlamaIndex 创始人 Jerry Liu 简短点评:模型上 Opus 5.5 稍好,客户端却是 Codex 更顺手,适合纠结选哪家的人参考
The handbook turns vague "AI native development" talk into concrete organisational and infrastructure decisions, from guardrails to permission systems, that teams rarely document publicly. Engineering managers and platform teams designing their own agent-based delivery pipelines can benchmark their setup against Alibaba's five-challenge framework. The full text is free to read online, and the MCP and CLI tooling sections align with workflows developers in the region already use.
一个叫 Relate 的开源小工具,帮你记下答应过别人的事和没回的消息,Claude 和 Cursor 都能读,朋友圈健忘症患者可以试试。
Alibaba 的 Wan3.0 工作流实测,10 秒 1080p 视频在 AskVenice 上只花 1.82 美元,算成本账的朋友可以看看。
Because the diagrams are plain text, agents can re-edit them later just as they wrote them, with no renderer or binary assets involved. Developers writing postmortems, refactor plans or PR notes can add the skill, request a write-up, and get a chart picked by the decision table in seconds. Compared with Mermaid or Recharts, this stays legible in any Markdown host, though the strict rules — max two charts per piece, no pie charts, Geist Mono only — demand discipline.
藏师傅拿 Opus 5.5 给 CodePilot 一键做了条宣传片,他看完直接说吊打 GPT-6 Astra,视频效果可以自己看看。
歸藏用 Opus 5.5 读 Zed 代码库做了条宣传视频,镜头跟着功能走还自动卡点,这条 Skill 思路可以抄。
有人用 Claude Code 挂 Opus 5.5,不写脚本不套模板,直接让模型从零做出一部《崂山道士》3D 片,思路很值得参考。
The takeaway cuts against the hype narrative: stronger models raise the skill floor for engineering judgement rather than removing the need for engineers. Developers should treat agents as tools requiring review discipline, not a replacement for software craft, and skeptics of 'everyone can code now' claims get a concrete counterpoint from two prominent practitioners. The exchange centres on Claude Opus 4.6 and GPT-5.2, the benchmarks many teams actually evaluate agents against.
有人用 ChatGPT 一个冷门新功能核对了 GPU 功率密度数据,960 机柜 200kW,但比 Blackwell 还差 4 倍。
Threshold selection, not the raw score, determines how much extraction work a pipeline can safely automate at a given precision target. Teams building document-processing agents can use the precision-recall trade-off framing to set human review volume. ExtractBench numbers offer a concrete way to compare LlamaParse Agentic Plus against other extraction systems before committing to one.
有人拿 Claude Opus 5.5 做了条《权游》私生子之战解说,配音还是克隆自己的声音,流程挺有意思,可以去看看成品。
Databricks 出了个可直接导入的 notebook,教你把 Jev 决策模型跑在自家 Serverless GPU 上,SQL 里调 ai_query 就能拿分类和概率分数,换上自己数据就能用。
想考 PyTorch 认证的话可以看下,10 月 19 日在圣何塞有一整天实操培训,还附赠考试券,10 月 13 日前得报名。
Virgin Voyages 把营销全面交给 AI 重做,直接看营收,结果拿下史上最强预订季,案例细节值得扒。
Claude Code 团队要把 plan mode 改成内置 mod,还能让 mods 自定义 shift+tab 模式切换,重度用户可以关注这个改动。
LangChain 把 LangSmith 里的真实运行轨迹直接变成微调数据,用 Baseten 和 Fireworks AI 训开源模型,做智能体的朋友可以抄这套流程。
做文档抽取的可以看看,LlamaIndex 用 ExtractBench 实测了置信度过滤怎么用,97% 精确率下 LlamaParse 召回 66.48%,教你算该放多少单自动跑。
有博主用 Opus 5.5 直接做产品视频,说镜头和文案比 GPT-6 还强,连 Computer Use 都不用,明天还有实操案例。
Together AI 手把手教你花 17 美元、25 分钟,基于 Qwen3.5 4B 微调出一个小模型,教程免费,想从零动手训模型的可以照着跑一遍。
有人把 Seedance 2.5 的人物替换玩法摸透了,还附带完整提示词,连被拒审的绕过办法都写了,做视频的可以直接抄作业。
GitHub 官方出了个教程,用 Copilot 自动把一堆 Dependabot 依赖 PR 按风险分组、查 CI、写摘要,每天上班前自动跑一遍,维护老项目的人可以试试。
Milvus 放了个电商搜索 demo,语义召回后直接在库里用 XGBoost 重排,不用再搭一层排序服务,代码也开源了。
一个编辑每周要读几百篇论文,他把自己的筛选流程拆成 12 条规则交给 agent 跑,规则全贴出来了,照抄就能用。
阿里云在云栖大会上讲怎么把大模型从聊天变成干活的生产力,拆成上下文、推理、规划、执行四步,做 Agent 的可以看看。
Lovable 创始人在播客里现场教学,从重设计官网到自动筛选合作邮件,一条工作流全跑通,还做了测试和安全扫描。
Only the latest 2000 items are listed; for older ones see the daily archive